基于三轴加速度传感器的人体行为识别研究

发布时间：2018-03-02 15:15

本文选题：加速度传感器　切入点：人体行为识别　出处：《江南大学》2014年硕士论文　论文类型：学位论文

【摘要】：基于加速度传感器的人体行为识别是模式识别领域中的一个新兴的研究方向，它的迅速发展受惠于微电子和传感器技术的不断进步以及模式识别理论的深入研究。随着人们对智能交互和健康监护等方面需求的日益增长，基于加速度传感器的人体行为识别在医疗保健、运动检测、能耗评估等领域受到了广泛的关注。与基于计算机视觉的行为识别不同，基于加速度传感器的方法更能体现人体运动的本质，而且不受特定的场景和时间限制，能量消耗少，成本较低，更适合推广应用。虽然近年来基于加速度传感器的行为识别取得了极大的进展，但仍面临不少急需解决的问题，包括如何提取具有较强表征能力的信号特征，如何面向实际应用设计合理的跌倒识别方法，如何构建高精度、泛化能力强的行为分类器等问题。围绕这些问题，本文主要进行了如下的研究工作： 1)总结了现有的行为识别方法，比较了基于计算机视觉和基于加速度传感器两种方法，详细分析了基于加速度信号的行为识别具有的优势，系统研究了该类方法的实现过程和相关技术。 2)针对行为识别过程中的特征提取问题，从加速度信号的时频分析和分布特点的角度出发，利用小波分析等技术手段，提取了基于角度的小波能量和关键点连线斜率两种新颖特征，从不同方面对加速度信号进行刻画。利用独立检测法和交叉验证法对不同特征集合的识别率进行了比较，，表明了这两种特征的有效性。 3)在跌倒识别方面，常用分类器往往需要大量的训练样本，现有的方法常采用故意反复跌倒的方式获取训练样本，但对于用户而言非常不便。针对这一问题，提出了一种基于隐马尔科夫模型和身体倾角的跌倒识别方法。该方法将跌倒识别问题转换为对已学模型的偏差问题进行处理，减小跌倒样本量对识别结果的影响。而且基于时序分析的方法，可以有效保留研究对象前后的状态信息，更加符合物理规律。 4)在日常行为识别方面，为了提高分类器的泛化能力和识别正确率，采用递阶遗传算法训练RBF神经网络，对其结构和参数同时寻优。以降低分类器结构复杂度和提高正确率为目的，设计了新的适应度函数，利用四分位间距改进参数基因的交叉方式，并结合两种变异操作，提高寻优效率。实验结果表明，采用改进递阶遗传算法训练的RBF网络分类器，同时具备结构精简和误差较低的优点，对7种行为的识别率可达91.54%。 5)从识别系统的底层出发，设计了一种加速度信号采集平台，实现了对运动加速度数据的采集。
[Abstract]:Human behavior recognition based on acceleration sensor is a new research direction in the field of pattern recognition. Its rapid development has benefited from the continuous progress of microelectronics and sensor technology, as well as the in-depth study of pattern recognition theory. With the increasing demand for intelligent interaction and health monitoring, Human behavior recognition based on accelerometer has attracted wide attention in the fields of medical care, motion detection, energy consumption evaluation, etc., which is different from behavior recognition based on computer vision. The method based on acceleration sensor can reflect the essence of human motion, and it is not limited by the specific scene and time, and the energy consumption is less and the cost is lower, so it is more suitable for popularization and application. Although great progress has been made in behavior recognition based on acceleration sensors in recent years, there are still many problems that need to be solved, including how to extract signal features with strong representation ability. How to design a reasonable fall recognition method for practical application and how to construct a high precision and strong generalization behavior classifier. Around these problems, this paper mainly carried out the following research work:. 1) summarizing the existing behavior recognition methods, comparing the two methods based on computer vision and acceleration sensor, and analyzing the advantages of behavior recognition based on acceleration signal in detail. The realization process and related technology of this kind of method are studied systematically. 2) aiming at the problem of feature extraction in the process of behavior recognition, from the point of view of time-frequency analysis and distribution characteristics of acceleration signal, wavelet analysis and other technical means are used. Two novel features of wavelet energy based on angle and slope of key points are extracted and the acceleration signals are depicted from different aspects. The recognition rates of different feature sets are compared by using independent detection method and cross-validation method. The validity of these two features is demonstrated. 3) in the aspect of fall recognition, the common classifier often needs a large number of training samples. The existing methods often use the method of intentional repeated fall to obtain the training sample, but it is very inconvenient for the user. A fall recognition method based on Hidden Markov Model (hmm) and body inclination angle is proposed in this paper. Based on the method of time series analysis, the state information before and after the study object can be effectively retained, which is more in line with the physical law. 4) in the aspect of daily behavior recognition, in order to improve the generalization ability and recognition accuracy of classifier, hierarchical genetic algorithm is used to train RBF neural network. In order to reduce the structural complexity and improve the accuracy of the classifier, a new fitness function is designed. The crossover of parameter genes is improved by using quartile spacing, and two mutation operations are combined. The experimental results show that the RBF classifier trained by improved hierarchical genetic algorithm has the advantages of simple structure and low error, and the recognition rate of seven behaviors can reach 91.54. 5) from the bottom of the recognition system, an acceleration signal acquisition platform is designed to collect the acceleration data.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP212;TN911.7

【参考文献】